The proliferation of malicious online behaviors, particularly cyber-trolling, presents significant challenges to maintaining healthy online communities. This paper investigates the efficacy of four deep learning architectures-BERT, LSTM, GRU, and Causal Convolutional Networks (Causal Conv 1D)-for the automatic detection of cyber-trolls based on textual content. Using a comprehensive dataset of 50,000 social media comments, we evaluate these models on their ability to distinguish between normal users and trolls. Our results indicate that while the pre-trained BERT model achieves the highest overall accuracy ($94.2 \%$), the Causal Conv 1D architecture demonstrates competitive performance $(92.7 \%)$ with significantly lower computational requirements. We also analyze the semantic features that most effectively contribute to troll detection and discuss the ethical implications of automated moderation systems. This research contributes to the development of more efficient and effective methods for maintaining civil discourse in online spaces.